Warwick Medical School, University of Warwick, Coventry, Warwickshire, UK.
Med Decis Making. 2021 Jan;41(1):37-50. doi: 10.1177/0272989X20973201. Epub 2020 Dec 7.
Extrapolations of parametric survival models fitted to censored data are routinely used in the assessment of health technologies to estimate mean survival, particularly in diseases that potentially reduce the life expectancy of patients. Akaike's information criterion (AIC) and Bayesian information criterion (BIC) are commonly used in health technology assessment alongside an assessment of plausibility to determine which statistical model best fits the data and should be used for prediction of long-term treatment effects. We compare fit and estimates of restricted mean survival time (RMST) from 8 parametric models and contrast models preferred in terms of AIC, BIC, and log-likelihood, without considering model plausibility. We assess the methods' suitability for selecting a parametric model through simulation of data replicating the follow-up of intervention arms for various time-to-event outcomes from 4 clinical trials. Follow-up was replicated through the consideration of recruitment duration and minimum and maximum follow-up times. Ten thousand simulations of each scenario were performed. We demonstrate that the different methods can result in disagreement over the best model and that it is inappropriate to base model selection solely on goodness-of-fit statistics without consideration of hazard behavior and plausibility of extrapolations. We show that typical trial follow-up can be unsuitable for extrapolation, resulting in unreliable estimation of multiple parameter models, and infer that selecting survival models based only on goodness-of-fit statistics is unsuitable due to the high level of uncertainty in a cost-effectiveness analysis. This article demonstrates the potential problems of overreliance on goodness-of-fit statistics when selecting a model for extrapolation. When follow-up is more mature, BIC appears superior to the other selection methods, selecting models with the most accurate and least biased estimates of RMST.
对删失数据拟合的参数生存模型的外推法通常用于评估卫生技术，以估计平均生存时间，特别是在那些可能降低患者预期寿命的疾病中。Akaike 信息准则（AIC）和贝叶斯信息准则（BIC）常用于卫生技术评估，同时评估模型的合理性，以确定哪个统计模型最适合数据，并应用于预测长期治疗效果。我们比较了 8 种参数模型的拟合度和限制性平均生存时间（RMST）的估计值，并根据 AIC、BIC 和对数似然对比了首选模型，而不考虑模型的合理性。我们通过模拟 4 项临床试验中各种时间事件结局的干预组随访数据，评估了这些方法在选择参数模型方面的适用性。通过考虑招募持续时间和最小和最大随访时间来复制随访。对每个场景进行了 10000 次模拟。我们证明，不同的方法可能导致对最佳模型的不一致意见，并且不考虑危险行为和外推的合理性，仅基于拟合优度统计数据来选择模型是不适当的。我们表明，典型的试验随访可能不适合外推，导致多个参数模型的估计不可靠，并推断由于成本效益分析中存在高度不确定性，仅基于拟合优度统计数据选择生存模型是不合适的。本文演示了当选择用于外推的模型时，过度依赖拟合优度统计数据的潜在问题。当随访更加成熟时，BIC 似乎优于其他选择方法，选择 RMST 估计最准确、偏差最小的模型。
